In light of the significance of wind power within the energy landscape and the challenges posed by its intermittency, this paper proposes an end-to-end, ultra-short-term wind power multi-step prediction model that integrates outlier processing and multi-scale feature fusion. The objective is to enhance the accuracy and stability of ultra-short-term wind power predictions, thereby providing robust support for the reliability of power system scheduling and operation. First, the RobustTSF method is employed to address time series anomalies, providing a strong assurance of the prediction model′s robustness and minimizing the disparity between abnormal time series prediction and noise label learning. Secondly, the integration of the spatial pyramid matching mapping strategy, Levy flight strategy, and adaptive T-distribution mutation strategy enhances the dung beetle optimization algorithm, significantly improving its global search capability and convergence efficiency. Meanwhile, the multi-strategy dung beetle optimization algorithm is utilized to optimize the hyperparameters of the enhanced TimeMixer model, resulting in optimal model performance. Finally, the CATimeMixer model is employed to achieve the fusion and prediction of multi-scale seasonal features and trend features. The experimental results indicate that the MAE, RMSE, and MSE decreased by 49.71%, 41.26%, and 65.50%, respectively, compared to the benchmark model multilayer perceptron, while the R2 value increased by 4.49%. This demonstrates a significant reduction in prediction error and offers a novel approach for the accurate prediction of ultra-short-term wind power.